A rest-trained SEEG decoder times mouse-movement onset better than a shifted null, but often crosses early

We evaluated whether a movement-versus-rest decoder could estimate mouse-movement onset on held-out single trials from archived 70–150 Hz SEEG features for seven patients in OpenNeuro ds006253. A shrinkage linear discriminant was trained on log-transformed pre-stimulus rest and movement windows. On each test trial, onset was the centre of the first 0.2 s bin above a threshold selected on training trials; the comparison null circularly shifted each test trajectory. Mean absolute timing error was 0.540 s for real trajectories and 0.662 s for shifted trajectories (all seven patients lower error; exact one-sided sign-flip p = 0.0078). The fraction of trials within 0.2 s of onset was 0.23 versus 0.11 (all seven patients higher; Holm-adjusted p = 0.031). However, 53% of trials crossed at least 0.3 s before onset and 12% had no crossing. Early crossings were less frequent in short reaction-time than long reaction-time trials (0.47 versus 0.58; none of seven patients higher in short-RT; Holm-adjusted p = 0.031), opposite to the pre-registered expectation that stimulus responses caused early crossings. The detector carried onset-time information, but its timing was coarse and often early. The study is limited to seven patients, one dataset, one time-ordered split, offline analysis, 0.2 s bins, and one threshold rule. The 1 s pre-onset span truncates very early crossings at -0.9 s; the reaction-time split mixes stimulus timing and deliberation, and this difference was not tested mechanistically.

SEEG · movement onset · brain-computer interface · shrinkage LDA · single-trial decoding · high-gamma activity · timing error · OpenNeuro ds006253